Claude Code Digest — 2026-05-16 00:33:32
Version updates
Version 2.1.143 introduces rigid plugin dependency enforcement. The claude plugin disable command now aborts if another active plugin requires your target. The CLI returns a copy-pasteable command chain to disable all dependent plugins instead. Conversely, claude plugin enable automatically provisions missing dependencies. This guarantees a stable toolchain and prevents isolated tool failures during complex multi-step tasks.
The release also introduces projected context costs. Claude Code now calculates anticipated token consumption per-turn and per-invocation before execution. Developers gain critical cost visibility upfront, allowing them to stop expensive runaway feedback loops.
What the docs reveal
Context Optimization via Dynamic Tool Loading
Loading dozens of Model Context Protocol (MCP) tools simultaneously bloats the context window. Anthropic solves this with the ENABLE_TOOL_SEARCH environment variable. You can now defer tool loading until the model specifically requests them.
Setting this variable to auto:N triggers dynamic discovery only when loaded tool definitions consume a specific percentage of the maximum context window. This prevents massive enterprise tool schemas from suffocating the prompt. Anthropic intentionally disabled this feature on Vertex AI and third-party proxies by default. Dynamic tool discovery requires specific hardware and newer API features (Claude 4.5+), so this default protects users from silent compatibility failures.
Execution Telemetry and Hook Data
The PostToolUse hook now emits detailed execution telemetry for subagents. Developers gain programmatic access to granular metrics: totalTokens, totalDurationMs, and detailed cache performance counts. You no longer need external proxy systems to calculate agent execution costs.
Asynchronous tasks behave differently. Subagent calls launched with run_in_background: true skip usage metrics entirely. They return a status: "async_launched" payload containing the agentId and the output file path. This separation keeps the main thread clean while allowing separate chron-style scripts to harvest the delayed results.
Enterprise Fleet Auditing
Anthropic continues adapting Claude Code for corporate IT management. The /status diagnostic command now traces active configuration layers to their source. It identifies whether a setting originated from an enterprise-managed plist, a Windows registry key, or a local file. The command also flags invalid JSON syntax in local configurations.
To prevent corruption during automated updates across a fleet, the CLI now creates timestamped backups of ~/.claude.json in a dedicated backups/ directory before triggering any migration scripts.
Provider-Specific Stability Fixes
Background operations like generating session titles previously defaulted to the fast and inexpensive Haiku model. Anthropic observed that developers using Amazon Bedrock, Google Vertex AI, or Microsoft Foundry frequently lack guaranteed Haiku availability across all corporate regions.
Claude Code now forces background tasks to use your primary selected model on these third-party platforms. You can restore the cheaper behavior by explicitly defining ANTHROPIC_DEFAULT_HAIKU_MODEL with an available regional model identifier.
Localized Privacy Controls
Running the /feedback command on third-party providers previously risked transmitting proprietary prompt transcripts outside the organization's certified cloud boundary. Anthropic halted this behavior.
The CLI now strictly redacts API keys and stores feedback locally in ~/.claude/feedback-bundles/ whenever you use Vertex, Bedrock, or Foundry. You retrieve these encrypted bundles and hand them to your Anthropic enterprise representative directly. To lock down endpoints entirely, administrators can set DISABLE_FEEDBACK_COMMAND to strip the feature for all users.